Breast cancer is the most prevalent genetic malignancy, causing cancer-related fatalities among women worldwide. This highlights the importance of early detection and accurate response evaluation in improving patient prognosis and long-term survival rates. Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCEMRI) captures voxel-wise temporal insights that reveal perfusion heterogeneity and vascular permeability, thereby facilitating the discrimination of pCR (pathological complete response) and non-pCR (non-pathological complete response) when evaluating treatment response. Accordingly, this work proposes a temporally derived fractional order derivative (FOD) based radiomics model for breast cancer treatment response classification. Herein, the model is primarily validated using the QIN Breast DCEMRI dataset, consisting of a limited patients (pCR=3, nonpCR=7) scanned at 32 time points. However, each patient encompasses prominent temporal insights, assisting in validating the model. Initially, the I-max maps are extracted from 4D DCE-MRIs, which are utilized for computing the Fractional Order Derivative (FOD) gradient images. A comprehensive set of radiomic features was extracted from the FOD gradient images corresponding to two chemotherapy cycles (Visit 1 and Visit 2). Generally, the FOD images provide complementary information, enhancing the predictive ability of the radiomics model. Further, a robust feature dimensionality reduction technique is incorporated for selecting optimal features from the combined feature sets (Visit 1, Visit 2, and.(V2, V1). The selected optimal features are utilized for training and testing the classification models, namely, Support Vector Machine (SVM) kernels (Linear, RBF, polynomial), Logistic Regression, and Random Forest for classifying pCR and non-pCR classes. Further, the classifying models are evaluated by metrics, obtaining an interpretable and balanced model. Hence, the clinical application of the proposed model aids personalised therapy and improves the survival rate.
Electroencephalogram (EEG) signals are often contaminated by cardiac artifacts, which can significantly degrade the performance of downstream analysis. This paper presents a Fourier Domain Denoising Convolutional Neural Network (FD-DCNN) that leverages frequency domain convolution to effectively suppress electrocardiogram (ECG) artifacts in single channel EEG signals. Unlike conventional, time domain CNN-based approaches, the proposed model operates entirely in the spectral domain, replacing computationally expensive convolutions with element-wise complex multiplications, suitable for area and power constrained biomedical hardware. A complex-domain non-linear activation, PhaseReLU, is introduced after each spectral convolution to preserve magnitude while enabling non-linear feature extraction. The network is trained using a Morphology Preserving Loss (MPL) combining mean squared error and correlation to maintain amplitude fidelity and waveform morphology. The publicly available EEGdenoiseNet and MIT-BIH Polysomnographic databases were used to generate the simulated dataset for model training and evaluation. FD-DCNN achieves a lower Relative Root Mean Square Error (RRMSE) of 0.4301 in the time domain and 0.3548 in the frequency domain, along with a higher correlation coefficient of 0.8879, while requiring only 0.67M parameters, 2.5MB of memory, and 2.31M floating point operations per second (FLOPs), which is substantially less than a comparable time domain CNN with similar architecture. These results establish FD-DCNN as a lightweight, high-performance denoising architecture well-suited for efficient hardware implementation and real-time EEG processing.
Seismocardiographic signal (SCG) is the chest vibration induced during the cardiac cycle due to cardiac muscle contraction, cardiac valves movement and blood momentum changes [ 1 – 5 ]. SCG can be used for estimating cardiac time intervals (CTI), which has been shown to be of potential value in predicting certain heart diseases [7 , 8] . Detecting CTI relies on accurate annotation of SCG and electrocardiogram (ECG) fiducial points including aortic valve opening (AO) and closure (AC), mitral valve opening (MO) and closure (MC) and ECG Q, R and T peaks [3 , 9] . Several studies addressed fiducial points detection using different methods [3 , 9 – 11] , however, a fully automated method that is simple and accurate is not yet available. Therefore, the objective of the current study is to introduce such a method and validate it by comparing the extracted CTI to the literature and highlight discrepancies in reporting CTI. The focus will be on the MC and AC events in SCG which are associated with the first and second heart sounds respectively [13] .
This review briefly investigates how artificial intelligence (AI) methods are being applied to microbiota-based clinical tasks. We analyse a curated set of peer-reviewed studies to characterize the models used, the nature of microbiota-derived inputs, and the clinical goals addressed. Our findings show a preference for classical machine learning approaches, especially random forests, due to their robustness and interpretability. Deep learning methods are less frequent and primarily employed in multimodal contexts. Most studies focus on disease prediction or classification, though some explore treatment response or drug-microbiota interactions. Gut-derived profiles dominate the input data, with limited exploration of other microbiota niches. Key challenges include the lack of external validation, inconsistent preprocessing practices, and limited use of explainability techniques. These observations point to the need for more standardized, transparent, and clinically grounded research to advance the integration of AI with microbiome science.
A non-invasive brain-computer interface (BCI) system based on electroencephalogram (EEG) signals is presented to facilitate intuitive control of assistive devices for individuals with motor impairments. EEG recordings corresponding to actual and imagined motor movements are transformed into two-dimensional representations us-ing the Gramian Angular Summation Field (GASF) technique, effectively encoding temporally significant fea- tures. The resulting GASF images are classified using the GoogLeNet convolutional neural network architecture, achieving a classification accuracy of up to 94.4% across both real and imagined movement tasks. Experimental results demonstrate that the proposed model attains higher accuracy in recognizing actual movements compared to imagery-based movements, indicating its potential as a practical solution for reliable motor intention decoding in assistive and rehabilitative applications.
The premature infant is prone to respiratory issues due to poor development of the regulatory system as well as mechanical issues with the lung itself. These problems often lead to apnea, the cessation of breathing. Until the respiratory system develops, the infant must be kept in critical care.
Crystallization processes in soft tissues have long been studied in relation to pathology and have been shown to positively correlate with precursory pathological cell activity. Early detection of breast cancer remains a daunting challenge in artificial intelligence, partly due to the prevalence of overdiagnosis and a lack of transparency in experimental results. Among histopathological specimens, tissue calcifications, ranging from dystrophic hydroxyapatite deposits to idiopathic oxalate crystals and psammoma bodies, have long held diagnostic promise but remain underexploited by visual models.In this paper, we introduce a novel crystallization-focused approach to multi-classification tasks derived from the Fox Chase Cancer Center Breast Tissue Corpus (FCBR). We constructed an annotated subset comprising 439 patches categorized into Crystalline Non-Neoplastic (cnno: n = 51), Crystalline Ductal Carcinoma in Situ (cdcs: n = 168), and Crystalline Invasive Ductal Carcinoma (cidc: n = 220).Leveraging this subset, we conducted a baseline experiment comparing a simple Random Forest classifier trained on (1) a standard, non-crystallization dataset (1,850 patches) versus (2) an enriched dataset including the crystallization annotations (2,243 patches). Models were evaluated on held-out FCBR samples (18,224 patches) and externally validated on the Temple University Hospital Digital Pathology (TUDP) Corpus (46,666 patches). Incorporation of crystallization annotations more than doubled the overall accuracy: from 18.4% to 34.6% on FCBR and from 20.7% to 23.5% on TUBR. These improvements persisted under domain shift, indicating enhanced generalization and statistical significance.Our findings demonstrate that explicit modeling of microcalcification patterns provides biologically meaningful features that strengthen deep learning based breast cancer detection. We propose that further expansion of crystallization annotations, especially for underrepresented non-neoplastic cases, and integration into more complex architectures may drive additional gains in sensitivity and diagnostic granularity.
Cancer cachexia is a complex, multifactorial syndrome characterized by severe muscle wasting, weight loss, and systemic inflammation, significantly affecting patient prognosis and quality of life. Cachexia is estimated to occur in up to 80% of people with advanced cancer, depending on cancer type and response to cancer treatment. Although it is said to be the direct cause of 30% of cancer deaths, there are currently no effective treatments for cachexia. The clinical definition of the syndrome is also loosely defined as losing more than 5% of body weight in the past six to twelve months. Accurate classification of cancer cachexia stages is crucial for timely intervention and personalized treatment strategies. Hidden Markov Models (HMMs), with their ability to model temporal sequences and hidden states, offer a robust approach for classifying the stages of cancer cachexia. This study explores the application of HMMs in analyzing longitudinal BMI and blood lab values from the MSK-IMPACT cohort to identify and classify the different stages of cachexia in cancer patients. The results demonstrate that HMMs can effectively distinguish between early, intermediate, and late stages of cachexia, thereby offering valuable insights for clinicians to optimize treatment regimens and improve patient outcomes. This approach improves the precision of cachexia staging and contributes to the broader field of predictive modeling in computational oncology.
Imagined speech is defined as the internal simulation of speaking without producing audible sound [1] . Brain–computer interfaces (BCIs) that decode imagined speech into text promise a communication channel for individuals with severe speech impairments. While most efforts have targeted word or phoneme level classification using electroencephalography (EEG), magnetoencephalography (MEG) and functional near infrared spectroscopy (fNIRS) as modalities, the capacity to decode continuous, coherent, and contextually relevant imagined speech remains as an active area of research. This review examines literatures that focuses on neuro-cognitive basis of imagined speech, non-invasive neural acquisition modalities, surveys signal processing and decoding methodologies, and scrutinize fluency-specific challenges and metrics, outlining benchmarks, current limitations. While prior reviews have addressed word-level and phoneme-level classification [3 , 4] , in this review we focus on fluency-specific challenges.
Emotional experiences during music listening are supported by dynamic interactions between perceptual, cognitive, and physiological processes, yet the extent to which bodily responses reflect or predict music-induced emotions remains an open question. In this preliminary study, we aimed to investigate the relationship between music listening enjoyment and a variety of physiological responses, testing whether any subset of these measures can be used to predict music-induced emotions. We presented each participant with obscure instrumental music excerpts, recorded several physiological responses during music listening, and asked them to subjectively rate their experienced valence (positive to negative affect) and arousal (high to low energy). Using an exploratory structural equation model, we found that high-frequency heart rate variability (HF HRV), skin conductance levels (SCL), and respiration rate had small yet significant correlations with arousal, while HF HRV and the range of skin conductance responses (SCR) were correlated with valence. These findings indicate that physiological responses like SCR and respiration rate could potentially serve as an objective measure of music-induced emotions, though the low predictive power of the model necessitates more expansive confirmatory analyses. Future research should expand upon these findings through larger sample sizes, cluster models, and by more closely investigating the influence of internal state.
Ultrasound imaging has been widely used in medical diagnosis for its safety, non-invasive, real-time imaging and other advantages [1] . However, the influence of the resolution limit of the ultrasound image and the inherent side lobe artifact is difficult to obtain the diagnostic information of difficult cases through the B-mode image [ 2 – 3 ]. At present, methods to improve the image quality mainly include three aspects: transmission strategy, adaptive beamforming methods, and image post-processing.
Recent advances in multimodal large language models (MLLMs) have opened new possibilities for biomedical image interpretation without task-specific training. This study explores the zero-shot visual reasoning capabilities of a leading MLLM, the ChatGPT vision model, for two challenging biomedical image classification tasks: electroencephalogram (EEG) signal interpretation and digital pathology (DPATH) image diagnosis. In this work, datasets of single-frame and three-frame EEG images and breast cancer pathology patches were used to benchmark performance. We show that while zero-shot MLLMs lag specialized models in accuracy, ChatGPT’s vision model delivers moderate performance and meaningful explanations compared to popular supervised computer vision models (ViT, ResNet). We also apply parameter-efficient fine-tuning (PEFT) to an open-source MLLM (the Qwen model) to improve accuracy across both domains. We find that off-the-shelf ChatGPT (o3-minihigh) can serve as a strong baseline model for biomedical tasks, highlighting the potential for model adaptation through lightweight supervised fine-tuning. The integration of AI-generated reasoning can enhance explainability and decision-making in clinical contexts.
In this study, we developed and evaluated a real-time implementation of the multivariate auto-regressive independent component analysis (MVARICA) model for estimating effective connectivity using Partial Directed Coherence (PDC). Each step of the MVARICA pipeline was adapted for online processing, with a focus on optimizing key hyperparameters, specifically model order and the delta ridge penalty in real time. The performance of the online model was benchmarked against a gold-standard offline MVARICA implementation. Our real-time model achieved a Mean Absolute Error (MAE) of 0.070 (7% error), with 95% of the value falling within a 20% deviation from the offline reference. Errors varied across frequency bands: Delta (MAE = 11.5%), Theta (9.1%), Alpha (7.8%), Beta (5.5%), and Gamma (3.6%). The Pearson correlation across all frequency bands exceeded 0.744, indicating strong agreement between the online and offline models. The real-time model successfully captured the key connectivity patterns identified by the offline version, converged reliably on hyperparameter optimization, and operated within real-time constraints for a low number of channels. Specifically, latency remained under 100 ms for low channel inputs but increased to approximately 1 second for high-density configurations.
Developmental dysplasia of the hip (DDH) is a pediatric condition where the hip joint is improperly formed resulting in abnormalities in femoral head, acetabulum or both, which can lead to mild acetabular shallowness or complete dislocation [1] . DDH incidence may be as high as 4% to 6% in newborn infants [2] , [3] . Prompt treatment such as Pavlik Harness or casting may ensure normal hip development and prevent long term disability. The older the age at DDH presentation, the worse the outcomes after intervention. As the infant grows, the non-invasive treatments often become ineffective necessitating surgical interventions, with generally poorer outcomes [4] , [5] .
Biofilms in the wild are composed of diverse microorganisms from multiple domains and are widespread in natural, human, and engineered environments. Biofilm-related problems exist across major economic and societal sectors, costing billions of dollars each year in energy losses, equipment damage, product contamination, agricultural losses, and medical infections. Understanding biofilms is crucial, particularly in how their heterogeneous structures and community organization affect their resilience, pathogenicity, nutrient access, and susceptibility to antimicrobial therapies.Digital imaging has become an essential instrument in biofilm research because it provides 3D spatial information of living biofilms in real-time. Existing biofilm image analysis software is available: the venerable COMSTAT, Semiautomated microbial abundance and morphology (CMEIAS), Daime, Beers Analysis and the newfangled BiofilmQ. Fundamentally, these approaches seek to answer a very simple yet crucial question: Where is the biomass? The answer remains surprisingly elusive, especially when the biofilm is very sparse or very thick. The next question is: Where are the microbes in the biofilm? When the individual microbes are densely packed, these approaches, that rely on thresholding, fail spectacularly. Cell segmentation remains an active area of research, with existing approaches including BCM3D 2.0, StarDist OPP, DeepSeeded, and Cellpose–SAM.Other more challenging research questions address higher order biofilm characteristics such as identification of microbial species, biofilm morphology and subcommunity arrangement. These require new robust image analysis methods that are easily accessible, centralized, computationally efficient, sensitive, and reproducible. The lack of such tools has hindered the meaningful and statistically robust integration of image analysis into biofilm studies. The Biofilm Imaging Library will overcome these hurdles and lay the groundwork for an open-access, community-driven repository of standardized, high-quality biofilm images, metadata, and associated computational models, and accelerate the development of open-source easy-to-use computationally efficient tools aided by machine learning. More about this resource is available at www.biofilmimagelibrary.org.
The present study extends previous work investigating eye-movement control in a dynamic, timeconstrained task environment. Earlier studies showed that the predictability of environmental dynamics influenced fixation allocation and the initiation sites of smooth pursuits. In those experiments, however, eye movements were clustered solely based on whether a reference point fell within foveal or peripheral vision, which may have confounded the analyses. This study introduces a bottomup, data-driven clustering approach to identify fixation types across multiple spatial and temporal dimensions. Six participants steered a spaceship to avoid obstacles under varying levels of motor control uncertainty. We identified two fixation types: Type 0 fixations, which were longer and centrally located near the spaceship, and Type 1 fixations, which occurred farther from the agent and were directed toward more open areas of the screen. Linear mixed modeling revealed that with increasing input noise, Type 0 fixations became shorter and more focused, while Type 1 fixations became longer and shifted farther from nearby obstacles. These patterns suggest adaptive gaze strategies, with Type 1 fixations potentially supporting predictive tracking under high control uncertainty. Our findings provide further evidence of how eye movements flexibly support action in complex, real-time environments.
The electrocardiogram (ECG) is an inexpensive and widely available tool for cardiac assessment. Despite its standardized format and small file size, the high complexity and inter-individual variability of ECG signals (typically a 60,000-size vector with 12 leads at 500 Hz) make it challenging to use in deep learning models, especially when only small training datasets are available. This study addresses these challenges by exploring feature generation methods from representative beat ECGs, focusing on Principal Component Analysis (PCA) and Autoencoders to reduce data complexity. We introduce three novel Variational Autoencoder (VAE) variants-Stochastic Autoencoder (SAE), Annealed ss-VAE (Ass-VAE), and Cyclical ss-VAE (Css-VAE)-and compare their effectiveness in maintaining signal fidelity and enhancing downstream prediction tasks using a Light Gradient Boost Machine (LGBM). The Ass-VAE achieved superior signal reconstruction, reducing the mean absolute error (MAE) to 15.7 +/- 3.2 mu V, which is at the level of signal noise. Moreover, the SAE encodings, when combined with traditional ECG summary features, improved the prediction of reduced Left Ventricular Ejection Fraction (LVEF), achieving an holdout test set area under the receiver operating characteristic curve (AUROC) of 0.901 with a LGBM classifier. This performance nearly matches the 0.909 AUROC of state-of-the-art CNN model but requires significantly less computational resources. Further, the ECG feature extraction-LGBM pipeline avoids overfitting and retains predictive performance when trained with less data. Our findings demonstrate that these VAE encodings are not only effective in simplifying ECG data but also provide a practical solution for applying deep learning in contexts with limited-scale labeled training data.
Despite significant advances in deep learningbased sleep stage classification, the clinical adoption of automatic classification models remains slow. One key challenge is the lack of explainability, as many models function as black boxes with millions of parameters. In response, recent work has increasingly focussed on enhancing model explainability. This study contributes to these efforts by introducing an explainability tool for spectral processing of individual EEG channels. Specifically, this tools retrieves the filter spectrum of low-level convolutional feature extraction and compares it with the classificationrelevant spectral information in the data. We apply our tool on the EEGNet and MSA-CNN models using the ISRUC-S3 and Sleep-EDF-20 datasets. The tool reveals that spectral processing plays a significant role in the lower frequency bands. In addition, comparing the correlation between filter spectrum and data-derived spectral information with univariate performance indicates that the model naturally prioritises the most informative channels in a multimodal setting. We specify how these insights can be leveraged to enhance model performance.
Analysis of intraneuronal signaling networks using systems biology approaches provides insights on how various molecules within each neuron may affect the behavior of the neuron under normal and pathological conditions. However, memory formation, learning and cognition are sophisticated functions of the human brain that not only depend on intraneuronal molecules and systems, but also emerge from the collective behavior of neurons in complex neuronal networks and the interneuronal processes among them. Therefore, understanding psychiatric and mental disorders where learning, memory or cognition are impaired, requires a hybrid modeling approach where both intraneuronal and interneuronal processes are included. In this paper, a hybrid model is introduced where a hippocampal CA1 neuronal network is considered, together with an intraneuronal signaling network that regulates the CREB (cAMP Response Element-Binding) protein. CREB is a transcription factor that is highly involved in cognitive and executive function of the human brain. Upon using the hybrid intraneuronal/interneuronal model, together with fault diagnosis analysis, we determine how neuronal excitability in the context of a neuronal network is affected, when there is one faulty - mutated or dysfunctional - molecule, or two concurrently faulty molecules. The hybrid approach allows to classify molecules into different classes, depending on how much they affect a neuronal network, when they are faulty. This has important applications in target discovery, since analysis of the hybrid model reveals which molecules or pairs of molecules result in substantial deviation from the normal network behavior, when they are faulty. Such molecules may be considered as proper targets, to develop effective therapeutics.
This research focuses on utilizing convolutional neural networks (CNNs) to identify biomarkers for major depressive disorder (MDD) in children and adolescents, compared to age-matched healthy individuals. We analyzed resting-state, eyes-closed electroencephalography (EEG) data, pre-processed and segmented by frequency bands and regions of interest (ROI). Several resting-state functional connectivity (rsFC) measurements were computed using a multi-variate auto-regressive (MVAR) model. The best-performing CNN model was further analyzed to understand its decision-making process and to identify relevant biomarkers. Our approach achieved an F1-Score of 0.790 and a Matthews correlation coefficient (MCC) of 0.745 using the full-frequency partial directed coherence (ffPDC) rsFC measurement. Among the connectivity metrics, partial directed coherence (PDC) outperformed coherence, partial coherence, and the directed transfer function (DTF). Additionally, the full-frequency versions of PDC and DTF demonstrated better performance compared to their standard and variant forms. These results highlight the potential of CNN models and EEG-derived biomarkers in advancing the understanding and diagnosis of MDD in children and adolescents.
